Green IT is a Business Strategy

So where does that leave leaders who need AI to deliver outcomes without compounding the problem? The answer isn’t far-off or a visionary gamble — its about making deliberate, practical choices. It’s operational discipline.

There are a few actions any AI leader can take now.

First, time non‑urgent workloads for cleaner power so emissions fall without sacrificing results. Second, set lifecycle emissions budgets for models the way you set spend caps, making trade‑offs explicit up front. Third, make GreenOps a standing team sport—shared metrics, shared incentives, visible dashboards. Fourth, choose regions and architectures that ease water stress, not just latency. Fifth, make impact visible by default with sustainability tagging in pipelines and products, so carbon and water show up next to cost and performance.

The paradox doesn’t vanish completely by integrating these actions into a larger sustainability strategy, but it becomes manageable—and progress becomes measurable. If we treat carbon, water, and energy as first‑class metrics alongside latency and accuracy, we’ll ship systems that earn trust while sharpening performance and resilience.

Carbon-Aware Scheduling for Workloads

We need to look closer at how AI workloads are executed. From training complex neural networks to deploying real-time inference systems, these workloads demand significant computational resources. Every model iteration, every batch of data processed, and every inference call consumes energy—often at a scale that’s invisible to end users but deeply impactful in aggregate. This raises an essential question: how and when should we choose to run these workloads?  This can make a substantial difference for sustainability.

Leaders and practitioners should consider running AI workloads when the grid is greenest. Most AI workloads are scheduled based on availability of compute resources or business urgency. But what if we scheduled them based on carbon intensity? For ROI-focused leaders, carbon-aware scheduling is a cost-and-risk lever, not a trade-off. Proven implementations show that shifting flexible workloads to cleaner, often off-peak windows lowers marginal energy costs and demand charges while cutting emissions that may soon carry explicit costs.

For example, Google’s carbon-intelligent platform already shifts non-urgent jobs without violating service level objectives (SLO), even supporting demand response during grid stress. Microsoft’s Azure pilots demonstrate measurable Software Carbon Intensity reductions by time-shifting batch jobs using carbon-intensity signals.

CASPER is a scheduling system that redirects traffic from high carbon regions to low-carbon and regions and determines the right amount of servers to provision. This has achieved up to ~70% emissions reductions while meeting latency targets, validating feasibility for production-grade services.

Itemize Power Usage for Finance Teams

Encourage finance to treat carbon like currency, approve emissions like spend, and manage tradeoffs. There are strategies for reducing the carbon emissions associated with machine learning (ML) that involve setting limits for carbon emissions for different stages of the project (training, tuning, and serving ML models). In software development, “Pull Requets (PRs)” are used to propose changes to code base. PR templates can include estimates for metrics such as cost per run, energy consumption, carbon emissions, and water usage.

Tools like CodeCarbon estimate the carbon dioxide emissions produced by machine learning experiments, tracking power usage and mapping it to the real-time carbon intensity of the electricity grid. These are lines of code that can be added to track your metrics.

Default to GreenOps

GreenOps are cross-functional teams that take ownership of executing sustainability goals and achieving operational efficiencies, cost savings, and enhanced corporate social responsibility. Some areas where GreenOps can perform:

• Reducing energy use in data centers
• Scheduling compute-heavy tasks during off-peak hours and using energy-efficient algorithms
• Monitoring and managing carbon footprints
• Optimizing servers, storage, and network bandwidth
• Using tools to measure the environmental impact of code and set emissions budgets
• Supporting sustainable IT hardware and software sourcing, minimizing waste, and promoting recyclability
• Choosing cloud services that prioritize renewable energy, energy-efficient data centers, and carbon and water tracking

There are already great examples out there of where GreenOps has executed. For example, UPS teamed with Google Cloud to develop route optimization software, resulting in annual savings of $400 million and a reduction of fuel consumption by 10 million gallons. Microsoft’s GreenOps teams optimize energy and water use in their global data centers, using AI-driven workload scheduling, closed-loop cooling systems, and real-time carbon tracking to help the company pursue its carbon-negative goal by 2030.

Despite growing interest, GreenOps remains uncommon in practice—especially across traditional, asset-heavy industries where sustainability responsibilities are fragmented. But, when sustainability is everyone’s job, it becomes no one’s burden.

Code with Carbon and Water in Mind

Choose cloud regions with lower water stress. This industry talks a lot about carbon, but water is the real hidden cost of AI.

AI data centers use millions of gallons of water for cooling. Water-aware deployment means choosing regions with lower water stress for inference-heavy workloads.

In 2024, hyperscale water use hit headline scale—Google and Microsoft data centers alone drew nearly 6 billion gallons, roughly the annual irrigation of about 40 golf courses—while a single 100MW facility using evaporative cooling can consume 500–700 million gallons a year, comparable to the needs of tens of thousands of residents.

Cluster a few sites in one metro — and withdrawals can push into the tens of billions of gallons, competing with municipal and industrial users in water‑stressed regions.

The fix for this is operational, not theoretical:

• Treat gallons per kWh as a hard constraint with metrics for efficiency of your water cooling system. This measurement is known as water usage effectiveness (WUE). A lower WUE means using less water per unit of IT power.
• Favor regions with lower water stress, and source reclaimed or non‑potable water by default.
• Move dense racks to closed‑loop liquid cooling (direct‑to‑chip or immersion) so coolant circulates in sealed loops, virtually eliminating evaporative losses after the initial fill. Additionally, pair with dry coolers and free‑air economization where climate allows, and use hybrid misting only on the hottest hours.
• Codify “water-awareness” in contracts and dashboards: add “water‑aware” to region selection next to carbon and latency
• Set a blue budget with monthly reporting, and require closed‑loop readiness and reclaimed‑water supply in new builds and coolers.
  

Tagging Code, Tracking Impact

Make environmental impact visible in your codebase. What gets measured gets managed—and what gets tagged gets noticed.

Sustainability tagging involves adding metadata to models and pipelines that track energy, carbon, and water usage. Teams can create tagging systems in GitHub or GitLab, include sustainability notes in model cards, and use badges or dashboards to visualize impact across projects.

Open-source projects like the Sustainable AI Energy Consumption Prediction Model are leading the way. This builds awareness, encourages accountability, and helps teams learn from each other.

AI Intelligence Starts with Responsibility

The AI sustainability paradox isn’t going away. But we can choose how we respond to it. These five ideas—carbon-aware scheduling, emissions budgeting, GreenOps, water-aware deployment, and sustainability tagging—are ideas for AI leaders that are practical and ready to be put in-place today. If we collectively make it our problem, we can solve it together.

If your AI strategy doesn’t include sustainability, is it really intelligent?